prov/verbs: Replace __BITS_PER_LONG with LONG_WIDTH

[raffenet]

The checks for the verbs provider can pass on a FreeBSD system, but will fail to compile because asm/types.h is not available.

Updated description: 2024-08-01
The verbs provider uses asm/types.h for __BITS_PER_LONG, but that header
is limited to Linux systems. LONG_WIDTH is defined in C23 in
limits.h. Use it instead, and add logic to define it if not
available. This will allow the verbs provider to compile on FreeBSD
systems.

[j-xiong]

The header is included in prov/verbs/src/ofi_verbs.h. Can verbs be built on BSD by commenting out the line? If so, a better fix would be using a conditional to guard the #include line.

[raffenet]

It does compile if I comment out that header, but gives this warning about a missing define.

--- prov/verbs/src/src_libfabric_la-verbs_ep.lo ---
prov/verbs/src/verbs_ep.c: In function 'vrb_query_atomic':
prov/verbs/src/verbs_ep.c:2002:5: warning: "__BITS_PER_LONG" is not defined, evaluates to 0 [-Wundef]
 2002 | #if __BITS_PER_LONG == 64

[j-xiong]

Something like this should work:

#if HAVE_ASM_TYPES_H
#include <asm/types.h>
#else
#if defined(__x86_64__) && !defined(__ILP32__)
#define __BITS_PER_LONG 64
#else
#define __BITS_PER_LONG 32
#endif
#endif

where HAVE_ASM_TYPES_H needs to be defined by configure.

[raffenet]

Something like this should work:

#if HAVE_ASM_TYPES_H
#include <asm/types.h>
#else
#if defined(__x86_64__) && !defined(__ILP32__)
#define __BITS_PER_LONG 64
#else
#define __BITS_PER_LONG 32
#endif
#endif

where HAVE_ASM_TYPES_H needs to be defined by configure.

OK. I will work on it later this week.

[raffenet]

OK. I will work on it later this week.

BTW, it looks like C23 adds LONG_WIDTH in limits.h which we could also test for before using the fallback code? See https://en.cppreference.com/w/c/types/limits

[j-xiong]

@raffenet I think that's something worth experimenting with (looks like it's already supported by gcc). Just need to keep the logic simple.

[raffenet]

@raffenet I think that's something worth experimenting with (looks like it's already supported by gcc). Just need to keep the logic simple.

What do you think about this? Looks like LONG_BIT is a non-standard define and LONG_WIDTH is the new standard. I also added support for __aarch64__ so that it sets the right value on my laptop. It's kind of verbose, so maybe it belongs in an osd.h header rather than the provider?

#include <limits.h>
#if !defined(LONG_WIDTH)
#ifdef LONG_BIT
#define LONG_WIDTH LONG_BIT
#elif defined(HAVE_ASM_TYPES_H)
#include <asm/types.h>
#define LONG_WIDTH __BITS_PER_LONG
#elif defined(__x86_64__) || defined(__aarch64__)
#ifdef !defined(__ILP32__)
#define LONG_WIDTH 64
#else
#define LONG_WIDTH 32
#endif
#endif
#endif

[j-xiong]

@raffenet Thanks! This looks good to me. Let's keep it in the verbs provider for now since it's only used here.

[raffenet]

Is the AWS CI failure real or no?

[wenduwan]

@raffenet Likely noise. I'm restarting it.

[wenduwan]

bot:aws:retest